AAPG Middle East Geoscience Technology Workshop, Integrated Emerging Exploration Concepts

Datapages, Inc.Print this page

A NN Supported Seismic Workflow to Create New Exploration Concepts Offshore Bahrain


Past exploration concepts for the area were targeting four-way dip closures, but the focus is now shifting towards the search of more complex structures including stratigraphic traps of various age. With significantly broadened computing capacities the application of machine learning algorithms becomes a useful standard procedure in seismic interpretation as it allows to dig deeper into the data and extract more useful information. We are presenting an integrated workflow using various neural network tools which helped us to create new exploration concepts offshore Bahrain. The basis for the analysis was a thoroughly reprocessed 3D seismic survey. The processing workflow included an anisotropic PSTM and a PSDM volume. The dynamic range of the data was broadened to lower and higher frequencies by the application of a deghosting algorithm, which later on proved to be very valuable for the classification of lithotypes from seismic data. A second important outcome of the reprocessing was a detailed well calibrated velocity model, which was used for depth conversion and to create the low frequency trends for a simultaneous inversion of angle stacks. Using the existing well data, a laterally homogenized facies model was built including anhydrites, tight and porous dolomites, tight and porous limestones, marls, ordinary shales, organic shales and sandstones. The volumetric model was then used to create electrofacies logs using Paradigm’s NN Facimage engine and to calculate missing shear wave logs in older wells. For the prediction of lithotypes from seismic data three different algorithms were tested – a standard lithoseismic classification from acoustic impedance and vp/vs ratio, a standard neural network inversion using an MLP algorithm and Paradigm’s new rock type classification using a democratic neural network association (DNNA). For each method, a starting model was built from upscaled well data, which was then propagated into 3D using various seismic attributes. As a result, we were able to build probabilistic lithology cubes for five rock types (anhydrites, tight carbonates, porous carbonates, shales, sandstones), as well as cubes for continuous properties like Vshale and porosity. The visualization of these cubes resulted in a better understanding of the petroleum system and led to the delineation of a new trap type – compressional structures related to subtle faults, which improved the exploration potential of the area.